net/smc/smc_pnet.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  Generic netlink support functions to configure an SMC-R PNET table
   6  *
   7  *  Copyright IBM Corp. 2016
   8  *
   9  *  Author(s):  Thomas Richter <tmricht@linux.vnet.ibm.com>
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/list.h>
  14 #include <linux/ctype.h>
  15 #include <net/netlink.h>
  16 #include <net/genetlink.h>
  17
  18 #include <uapi/linux/if.h>
  19 #include <uapi/linux/smc.h>
  20
  21 #include <rdma/ib_verbs.h>
  22
  23 #include <net/netns/generic.h>
  24 #include "smc_netns.h"
  25
  26 #include "smc_pnet.h"
  27 #include "smc_ib.h"
  28 #include "smc_ism.h"
  29 #include "smc_core.h"
  30
  31 #define SMC_ASCII_BLANK 32
  32
  33 static struct net_device *pnet_find_base_ndev(struct net_device *ndev);
  34
  35 static const struct nla_policy smc_pnet_policy[SMC_PNETID_MAX + 1] = {
  36         [SMC_PNETID_NAME] = {
  37                 .type = NLA_NUL_STRING,
  38                 .len = SMC_MAX_PNETID_LEN
  39         },
  40         [SMC_PNETID_ETHNAME] = {
  41                 .type = NLA_NUL_STRING,
  42                 .len = IFNAMSIZ - 1
  43         },
  44         [SMC_PNETID_IBNAME] = {
  45                 .type = NLA_NUL_STRING,
  46                 .len = IB_DEVICE_NAME_MAX - 1
  47         },
  48         [SMC_PNETID_IBPORT] = { .type = NLA_U8 }
  49 };
  50
  51 static struct genl_family smc_pnet_nl_family;
  52
  53 enum smc_pnet_nametype {
  54         SMC_PNET_ETH    = 1,
  55         SMC_PNET_IB     = 2,
  56 };
  57
  58 /* pnet entry stored in pnet table */
  59 struct smc_pnetentry {
  60         struct list_head list;
  61         char pnet_name[SMC_MAX_PNETID_LEN + 1];
  62         enum smc_pnet_nametype type;
  63         union {
  64                 struct {
  65                         char eth_name[IFNAMSIZ + 1];
  66                         struct net_device *ndev;
  67                 };
  68                 struct {
  69                         char ib_name[IB_DEVICE_NAME_MAX + 1];
  70                         u8 ib_port;
  71                 };
  72         };
  73 };
  74
  75 /* Check if two given pnetids match */
  76 static bool smc_pnet_match(u8 *pnetid1, u8 *pnetid2)
  77 {
  78         int i;
  79
  80         for (i = 0; i < SMC_MAX_PNETID_LEN; i++) {
  81                 if ((pnetid1[i] == 0 || pnetid1[i] == SMC_ASCII_BLANK) &&
  82                     (pnetid2[i] == 0 || pnetid2[i] == SMC_ASCII_BLANK))
  83                         break;
  84                 if (pnetid1[i] != pnetid2[i])
  85                         return false;
  86         }
  87         return true;
  88 }
  89
  90 /* Remove a pnetid from the pnet table.
  91  */
  92 static int smc_pnet_remove_by_pnetid(struct net *net, char *pnet_name)
  93 {
  94         struct smc_pnetentry *pnetelem, *tmp_pe;
  95         struct smc_pnettable *pnettable;
  96         struct smc_ib_device *ibdev;
  97         struct smcd_dev *smcd_dev;
  98         struct smc_net *sn;
  99         int rc = -ENOENT;
 100         int ibport;
 101
 102         /* get pnettable for namespace */
 103         sn = net_generic(net, smc_net_id);
 104         pnettable = &sn->pnettable;
 105
 106         /* remove table entry */
 107         write_lock(&pnettable->lock);
 108         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist,
 109                                  list) {
 110                 if (!pnet_name ||
 111                     smc_pnet_match(pnetelem->pnet_name, pnet_name)) {
 112                         list_del(&pnetelem->list);
 113                         if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev) {
 114                                 dev_put(pnetelem->ndev);
 115                                 pr_warn_ratelimited("smc: net device %s "
 116                                                     "erased user defined "
 117                                                     "pnetid %.16s\n",
 118                                                     pnetelem->eth_name,
 119                                                     pnetelem->pnet_name);
 120                         }
 121                         kfree(pnetelem);
 122                         rc = 0;
 123                 }
 124         }
 125         write_unlock(&pnettable->lock);
 126
 127         /* if this is not the initial namespace, stop here */
 128         if (net != &init_net)
 129                 return rc;
 130
 131         /* remove ib devices */
 132         spin_lock(&smc_ib_devices.lock);
 133         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 134                 for (ibport = 0; ibport < SMC_MAX_PORTS; ibport++) {
 135                         if (ibdev->pnetid_by_user[ibport] &&
 136                             (!pnet_name ||
 137                              smc_pnet_match(pnet_name,
 138                                             ibdev->pnetid[ibport]))) {
 139                                 pr_warn_ratelimited("smc: ib device %s ibport "
 140                                                     "%d erased user defined "
 141                                                     "pnetid %.16s\n",
 142                                                     ibdev->ibdev->name,
 143                                                     ibport + 1,
 144                                                     ibdev->pnetid[ibport]);
 145                                 memset(ibdev->pnetid[ibport], 0,
 146                                        SMC_MAX_PNETID_LEN);
 147                                 ibdev->pnetid_by_user[ibport] = false;
 148                                 rc = 0;
 149                         }
 150                 }
 151         }
 152         spin_unlock(&smc_ib_devices.lock);
 153         /* remove smcd devices */
 154         spin_lock(&smcd_dev_list.lock);
 155         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 156                 if (smcd_dev->pnetid_by_user &&
 157                     (!pnet_name ||
 158                      smc_pnet_match(pnet_name, smcd_dev->pnetid))) {
 159                         pr_warn_ratelimited("smc: smcd device %s "
 160                                             "erased user defined pnetid "
 161                                             "%.16s\n", dev_name(&smcd_dev->dev),
 162                                             smcd_dev->pnetid);
 163                         memset(smcd_dev->pnetid, 0, SMC_MAX_PNETID_LEN);
 164                         smcd_dev->pnetid_by_user = false;
 165                         rc = 0;
 166                 }
 167         }
 168         spin_unlock(&smcd_dev_list.lock);
 169         return rc;
 170 }
 171
 172 /* Add the reference to a given network device to the pnet table.
 173  */
 174 static int smc_pnet_add_by_ndev(struct net_device *ndev)
 175 {
 176         struct smc_pnetentry *pnetelem, *tmp_pe;
 177         struct smc_pnettable *pnettable;
 178         struct net *net = dev_net(ndev);
 179         struct smc_net *sn;
 180         int rc = -ENOENT;
 181
 182         /* get pnettable for namespace */
 183         sn = net_generic(net, smc_net_id);
 184         pnettable = &sn->pnettable;
 185
 186         write_lock(&pnettable->lock);
 187         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
 188                 if (pnetelem->type == SMC_PNET_ETH && !pnetelem->ndev &&
 189                     !strncmp(pnetelem->eth_name, ndev->name, IFNAMSIZ)) {
 190                         dev_hold(ndev);
 191                         pnetelem->ndev = ndev;
 192                         rc = 0;
 193                         pr_warn_ratelimited("smc: adding net device %s with "
 194                                             "user defined pnetid %.16s\n",
 195                                             pnetelem->eth_name,
 196                                             pnetelem->pnet_name);
 197                         break;
 198                 }
 199         }
 200         write_unlock(&pnettable->lock);
 201         return rc;
 202 }
 203
 204 /* Remove the reference to a given network device from the pnet table.
 205  */
 206 static int smc_pnet_remove_by_ndev(struct net_device *ndev)
 207 {
 208         struct smc_pnetentry *pnetelem, *tmp_pe;
 209         struct smc_pnettable *pnettable;
 210         struct net *net = dev_net(ndev);
 211         struct smc_net *sn;
 212         int rc = -ENOENT;
 213
 214         /* get pnettable for namespace */
 215         sn = net_generic(net, smc_net_id);
 216         pnettable = &sn->pnettable;
 217
 218         write_lock(&pnettable->lock);
 219         list_for_each_entry_safe(pnetelem, tmp_pe, &pnettable->pnetlist, list) {
 220                 if (pnetelem->type == SMC_PNET_ETH && pnetelem->ndev == ndev) {
 221                         dev_put(pnetelem->ndev);
 222                         pnetelem->ndev = NULL;
 223                         rc = 0;
 224                         pr_warn_ratelimited("smc: removing net device %s with "
 225                                             "user defined pnetid %.16s\n",
 226                                             pnetelem->eth_name,
 227                                             pnetelem->pnet_name);
 228                         break;
 229                 }
 230         }
 231         write_unlock(&pnettable->lock);
 232         return rc;
 233 }
 234
 235 /* Apply pnetid to ib device when no pnetid is set.
 236  */
 237 static bool smc_pnet_apply_ib(struct smc_ib_device *ib_dev, u8 ib_port,
 238                               char *pnet_name)
 239 {
 240         u8 pnet_null[SMC_MAX_PNETID_LEN] = {0};
 241         bool applied = false;
 242
 243         spin_lock(&smc_ib_devices.lock);
 244         if (smc_pnet_match(ib_dev->pnetid[ib_port - 1], pnet_null)) {
 245                 memcpy(ib_dev->pnetid[ib_port - 1], pnet_name,
 246                        SMC_MAX_PNETID_LEN);
 247                 ib_dev->pnetid_by_user[ib_port - 1] = true;
 248                 applied = true;
 249         }
 250         spin_unlock(&smc_ib_devices.lock);
 251         return applied;
 252 }
 253
 254 /* Apply pnetid to smcd device when no pnetid is set.
 255  */
 256 static bool smc_pnet_apply_smcd(struct smcd_dev *smcd_dev, char *pnet_name)
 257 {
 258         u8 pnet_null[SMC_MAX_PNETID_LEN] = {0};
 259         bool applied = false;
 260
 261         spin_lock(&smcd_dev_list.lock);
 262         if (smc_pnet_match(smcd_dev->pnetid, pnet_null)) {
 263                 memcpy(smcd_dev->pnetid, pnet_name, SMC_MAX_PNETID_LEN);
 264                 smcd_dev->pnetid_by_user = true;
 265                 applied = true;
 266         }
 267         spin_unlock(&smcd_dev_list.lock);
 268         return applied;
 269 }
 270
 271 /* The limit for pnetid is 16 characters.
 272  * Valid characters should be (single-byte character set) a-z, A-Z, 0-9.
 273  * Lower case letters are converted to upper case.
 274  * Interior blanks should not be used.
 275  */
 276 static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
 277 {
 278         char *bf = skip_spaces(pnet_name);
 279         size_t len = strlen(bf);
 280         char *end = bf + len;
 281
 282         if (!len)
 283                 return false;
 284         while (--end >= bf && isspace(*end))
 285                 ;
 286         if (end - bf >= SMC_MAX_PNETID_LEN)
 287                 return false;
 288         while (bf <= end) {
 289                 if (!isalnum(*bf))
 290                         return false;
 291                 *pnetid++ = islower(*bf) ? toupper(*bf) : *bf;
 292                 bf++;
 293         }
 294         *pnetid = '\0';
 295         return true;
 296 }
 297
 298 /* Find an infiniband device by a given name. The device might not exist. */
 299 static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
 300 {
 301         struct smc_ib_device *ibdev;
 302
 303         spin_lock(&smc_ib_devices.lock);
 304         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 305                 if (!strncmp(ibdev->ibdev->name, ib_name,
 306                              sizeof(ibdev->ibdev->name)) ||
 307                     !strncmp(dev_name(ibdev->ibdev->dev.parent), ib_name,
 308                              IB_DEVICE_NAME_MAX - 1)) {
 309                         goto out;
 310                 }
 311         }
 312         ibdev = NULL;
 313 out:
 314         spin_unlock(&smc_ib_devices.lock);
 315         return ibdev;
 316 }
 317
 318 /* Find an smcd device by a given name. The device might not exist. */
 319 static struct smcd_dev *smc_pnet_find_smcd(char *smcd_name)
 320 {
 321         struct smcd_dev *smcd_dev;
 322
 323         spin_lock(&smcd_dev_list.lock);
 324         list_for_each_entry(smcd_dev, &smcd_dev_list.list, list) {
 325                 if (!strncmp(dev_name(&smcd_dev->dev), smcd_name,
 326                              IB_DEVICE_NAME_MAX - 1))
 327                         goto out;
 328         }
 329         smcd_dev = NULL;
 330 out:
 331         spin_unlock(&smcd_dev_list.lock);
 332         return smcd_dev;
 333 }
 334
 335 static int smc_pnet_add_eth(struct smc_pnettable *pnettable, struct net *net,
 336                             char *eth_name, char *pnet_name)
 337 {
 338         struct smc_pnetentry *tmp_pe, *new_pe;
 339         struct net_device *ndev, *base_ndev;
 340         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 341         bool new_netdev;
 342         int rc;
 343
 344         /* check if (base) netdev already has a pnetid. If there is one, we do
 345          * not want to add a pnet table entry
 346          */
 347         rc = -EEXIST;
 348         ndev = dev_get_by_name(net, eth_name);  /* dev_hold() */
 349         if (ndev) {
 350                 base_ndev = pnet_find_base_ndev(ndev);
 351                 if (!smc_pnetid_by_dev_port(base_ndev->dev.parent,
 352                                             base_ndev->dev_port, ndev_pnetid))
 353                         goto out_put;
 354         }
 355
 356         /* add a new netdev entry to the pnet table if there isn't one */
 357         rc = -ENOMEM;
 358         new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
 359         if (!new_pe)
 360                 goto out_put;
 361         new_pe->type = SMC_PNET_ETH;
 362         memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
 363         strncpy(new_pe->eth_name, eth_name, IFNAMSIZ);
 364         new_pe->ndev = ndev;
 365
 366         rc = -EEXIST;
 367         new_netdev = true;
 368         write_lock(&pnettable->lock);
 369         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 370                 if (tmp_pe->type == SMC_PNET_ETH &&
 371                     !strncmp(tmp_pe->eth_name, eth_name, IFNAMSIZ)) {
 372                         new_netdev = false;
 373                         break;
 374                 }
 375         }
 376         if (new_netdev) {
 377                 list_add_tail(&new_pe->list, &pnettable->pnetlist);
 378                 write_unlock(&pnettable->lock);
 379         } else {
 380                 write_unlock(&pnettable->lock);
 381                 kfree(new_pe);
 382                 goto out_put;
 383         }
 384         if (ndev)
 385                 pr_warn_ratelimited("smc: net device %s "
 386                                     "applied user defined pnetid %.16s\n",
 387                                     new_pe->eth_name, new_pe->pnet_name);
 388         return 0;
 389
 390 out_put:
 391         if (ndev)
 392                 dev_put(ndev);
 393         return rc;
 394 }
 395
 396 static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
 397                            u8 ib_port, char *pnet_name)
 398 {
 399         struct smc_pnetentry *tmp_pe, *new_pe;
 400         struct smc_ib_device *ib_dev;
 401         bool smcddev_applied = true;
 402         bool ibdev_applied = true;
 403         struct smcd_dev *smcd_dev;
 404         bool new_ibdev;
 405
 406         /* try to apply the pnetid to active devices */
 407         ib_dev = smc_pnet_find_ib(ib_name);
 408         if (ib_dev) {
 409                 ibdev_applied = smc_pnet_apply_ib(ib_dev, ib_port, pnet_name);
 410                 if (ibdev_applied)
 411                         pr_warn_ratelimited("smc: ib device %s ibport %d "
 412                                             "applied user defined pnetid "
 413                                             "%.16s\n", ib_dev->ibdev->name,
 414                                             ib_port,
 415                                             ib_dev->pnetid[ib_port - 1]);
 416         }
 417         smcd_dev = smc_pnet_find_smcd(ib_name);
 418         if (smcd_dev) {
 419                 smcddev_applied = smc_pnet_apply_smcd(smcd_dev, pnet_name);
 420                 if (smcddev_applied)
 421                         pr_warn_ratelimited("smc: smcd device %s "
 422                                             "applied user defined pnetid "
 423                                             "%.16s\n", dev_name(&smcd_dev->dev),
 424                                             smcd_dev->pnetid);
 425         }
 426         /* Apply fails when a device has a hardware-defined pnetid set, do not
 427          * add a pnet table entry in that case.
 428          */
 429         if (!ibdev_applied || !smcddev_applied)
 430                 return -EEXIST;
 431
 432         /* add a new ib entry to the pnet table if there isn't one */
 433         new_pe = kzalloc(sizeof(*new_pe), GFP_KERNEL);
 434         if (!new_pe)
 435                 return -ENOMEM;
 436         new_pe->type = SMC_PNET_IB;
 437         memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
 438         strncpy(new_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX);
 439         new_pe->ib_port = ib_port;
 440
 441         new_ibdev = true;
 442         write_lock(&pnettable->lock);
 443         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 444                 if (tmp_pe->type == SMC_PNET_IB &&
 445                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
 446                         new_ibdev = false;
 447                         break;
 448                 }
 449         }
 450         if (new_ibdev) {
 451                 list_add_tail(&new_pe->list, &pnettable->pnetlist);
 452                 write_unlock(&pnettable->lock);
 453         } else {
 454                 write_unlock(&pnettable->lock);
 455                 kfree(new_pe);
 456         }
 457         return (new_ibdev) ? 0 : -EEXIST;
 458 }
 459
 460 /* Append a pnetid to the end of the pnet table if not already on this list.
 461  */
 462 static int smc_pnet_enter(struct net *net, struct nlattr *tb[])
 463 {
 464         char pnet_name[SMC_MAX_PNETID_LEN + 1];
 465         struct smc_pnettable *pnettable;
 466         bool new_netdev = false;
 467         bool new_ibdev = false;
 468         struct smc_net *sn;
 469         u8 ibport = 1;
 470         char *string;
 471         int rc;
 472
 473         /* get pnettable for namespace */
 474         sn = net_generic(net, smc_net_id);
 475         pnettable = &sn->pnettable;
 476
 477         rc = -EINVAL;
 478         if (!tb[SMC_PNETID_NAME])
 479                 goto error;
 480         string = (char *)nla_data(tb[SMC_PNETID_NAME]);
 481         if (!smc_pnetid_valid(string, pnet_name))
 482                 goto error;
 483
 484         if (tb[SMC_PNETID_ETHNAME]) {
 485                 string = (char *)nla_data(tb[SMC_PNETID_ETHNAME]);
 486                 rc = smc_pnet_add_eth(pnettable, net, string, pnet_name);
 487                 if (!rc)
 488                         new_netdev = true;
 489                 else if (rc != -EEXIST)
 490                         goto error;
 491         }
 492
 493         /* if this is not the initial namespace, stop here */
 494         if (net != &init_net)
 495                 return new_netdev ? 0 : -EEXIST;
 496
 497         rc = -EINVAL;
 498         if (tb[SMC_PNETID_IBNAME]) {
 499                 string = (char *)nla_data(tb[SMC_PNETID_IBNAME]);
 500                 string = strim(string);
 501                 if (tb[SMC_PNETID_IBPORT]) {
 502                         ibport = nla_get_u8(tb[SMC_PNETID_IBPORT]);
 503                         if (ibport < 1 || ibport > SMC_MAX_PORTS)
 504                                 goto error;
 505                 }
 506                 rc = smc_pnet_add_ib(pnettable, string, ibport, pnet_name);
 507                 if (!rc)
 508                         new_ibdev = true;
 509                 else if (rc != -EEXIST)
 510                         goto error;
 511         }
 512         return (new_netdev || new_ibdev) ? 0 : -EEXIST;
 513
 514 error:
 515         return rc;
 516 }
 517
 518 /* Convert an smc_pnetentry to a netlink attribute sequence */
 519 static int smc_pnet_set_nla(struct sk_buff *msg,
 520                             struct smc_pnetentry *pnetelem)
 521 {
 522         if (nla_put_string(msg, SMC_PNETID_NAME, pnetelem->pnet_name))
 523                 return -1;
 524         if (pnetelem->type == SMC_PNET_ETH) {
 525                 if (nla_put_string(msg, SMC_PNETID_ETHNAME,
 526                                    pnetelem->eth_name))
 527                         return -1;
 528         } else {
 529                 if (nla_put_string(msg, SMC_PNETID_ETHNAME, "n/a"))
 530                         return -1;
 531         }
 532         if (pnetelem->type == SMC_PNET_IB) {
 533                 if (nla_put_string(msg, SMC_PNETID_IBNAME, pnetelem->ib_name) ||
 534                     nla_put_u8(msg, SMC_PNETID_IBPORT, pnetelem->ib_port))
 535                         return -1;
 536         } else {
 537                 if (nla_put_string(msg, SMC_PNETID_IBNAME, "n/a") ||
 538                     nla_put_u8(msg, SMC_PNETID_IBPORT, 0xff))
 539                         return -1;
 540         }
 541
 542         return 0;
 543 }
 544
 545 static int smc_pnet_add(struct sk_buff *skb, struct genl_info *info)
 546 {
 547         struct net *net = genl_info_net(info);
 548
 549         return smc_pnet_enter(net, info->attrs);
 550 }
 551
 552 static int smc_pnet_del(struct sk_buff *skb, struct genl_info *info)
 553 {
 554         struct net *net = genl_info_net(info);
 555
 556         if (!info->attrs[SMC_PNETID_NAME])
 557                 return -EINVAL;
 558         return smc_pnet_remove_by_pnetid(net,
 559                                 (char *)nla_data(info->attrs[SMC_PNETID_NAME]));
 560 }
 561
 562 static int smc_pnet_dump_start(struct netlink_callback *cb)
 563 {
 564         cb->args[0] = 0;
 565         return 0;
 566 }
 567
 568 static int smc_pnet_dumpinfo(struct sk_buff *skb,
 569                              u32 portid, u32 seq, u32 flags,
 570                              struct smc_pnetentry *pnetelem)
 571 {
 572         void *hdr;
 573
 574         hdr = genlmsg_put(skb, portid, seq, &smc_pnet_nl_family,
 575                           flags, SMC_PNETID_GET);
 576         if (!hdr)
 577                 return -ENOMEM;
 578         if (smc_pnet_set_nla(skb, pnetelem) < 0) {
 579                 genlmsg_cancel(skb, hdr);
 580                 return -EMSGSIZE;
 581         }
 582         genlmsg_end(skb, hdr);
 583         return 0;
 584 }
 585
 586 static int _smc_pnet_dump(struct net *net, struct sk_buff *skb, u32 portid,
 587                           u32 seq, u8 *pnetid, int start_idx)
 588 {
 589         struct smc_pnettable *pnettable;
 590         struct smc_pnetentry *pnetelem;
 591         struct smc_net *sn;
 592         int idx = 0;
 593
 594         /* get pnettable for namespace */
 595         sn = net_generic(net, smc_net_id);
 596         pnettable = &sn->pnettable;
 597
 598         /* dump pnettable entries */
 599         read_lock(&pnettable->lock);
 600         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 601                 if (pnetid && !smc_pnet_match(pnetelem->pnet_name, pnetid))
 602                         continue;
 603                 if (idx++ < start_idx)
 604                         continue;
 605                 /* if this is not the initial namespace, dump only netdev */
 606                 if (net != &init_net && pnetelem->type != SMC_PNET_ETH)
 607                         continue;
 608                 if (smc_pnet_dumpinfo(skb, portid, seq, NLM_F_MULTI,
 609                                       pnetelem)) {
 610                         --idx;
 611                         break;
 612                 }
 613         }
 614         read_unlock(&pnettable->lock);
 615         return idx;
 616 }
 617
 618 static int smc_pnet_dump(struct sk_buff *skb, struct netlink_callback *cb)
 619 {
 620         struct net *net = sock_net(skb->sk);
 621         int idx;
 622
 623         idx = _smc_pnet_dump(net, skb, NETLINK_CB(cb->skb).portid,
 624                              cb->nlh->nlmsg_seq, NULL, cb->args[0]);
 625
 626         cb->args[0] = idx;
 627         return skb->len;
 628 }
 629
 630 /* Retrieve one PNETID entry */
 631 static int smc_pnet_get(struct sk_buff *skb, struct genl_info *info)
 632 {
 633         struct net *net = genl_info_net(info);
 634         struct sk_buff *msg;
 635         void *hdr;
 636
 637         if (!info->attrs[SMC_PNETID_NAME])
 638                 return -EINVAL;
 639
 640         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 641         if (!msg)
 642                 return -ENOMEM;
 643
 644         _smc_pnet_dump(net, msg, info->snd_portid, info->snd_seq,
 645                        nla_data(info->attrs[SMC_PNETID_NAME]), 0);
 646
 647         /* finish multi part message and send it */
 648         hdr = nlmsg_put(msg, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
 649                         NLM_F_MULTI);
 650         if (!hdr) {
 651                 nlmsg_free(msg);
 652                 return -EMSGSIZE;
 653         }
 654         return genlmsg_reply(msg, info);
 655 }
 656
 657 /* Remove and delete all pnetids from pnet table.
 658  */
 659 static int smc_pnet_flush(struct sk_buff *skb, struct genl_info *info)
 660 {
 661         struct net *net = genl_info_net(info);
 662
 663         smc_pnet_remove_by_pnetid(net, NULL);
 664         return 0;
 665 }
 666
 667 /* SMC_PNETID generic netlink operation definition */
 668 static const struct genl_ops smc_pnet_ops[] = {
 669         {
 670                 .cmd = SMC_PNETID_GET,
 671                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 672                 /* can be retrieved by unprivileged users */
 673                 .doit = smc_pnet_get,
 674                 .dumpit = smc_pnet_dump,
 675                 .start = smc_pnet_dump_start
 676         },
 677         {
 678                 .cmd = SMC_PNETID_ADD,
 679                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 680                 .flags = GENL_ADMIN_PERM,
 681                 .doit = smc_pnet_add
 682         },
 683         {
 684                 .cmd = SMC_PNETID_DEL,
 685                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 686                 .flags = GENL_ADMIN_PERM,
 687                 .doit = smc_pnet_del
 688         },
 689         {
 690                 .cmd = SMC_PNETID_FLUSH,
 691                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 692                 .flags = GENL_ADMIN_PERM,
 693                 .doit = smc_pnet_flush
 694         }
 695 };
 696
 697 /* SMC_PNETID family definition */
 698 static struct genl_family smc_pnet_nl_family __ro_after_init = {
 699         .hdrsize = 0,
 700         .name = SMCR_GENL_FAMILY_NAME,
 701         .version = SMCR_GENL_FAMILY_VERSION,
 702         .maxattr = SMC_PNETID_MAX,
 703         .policy = smc_pnet_policy,
 704         .netnsok = true,
 705         .module = THIS_MODULE,
 706         .ops = smc_pnet_ops,
 707         .n_ops =  ARRAY_SIZE(smc_pnet_ops)
 708 };
 709
 710 static int smc_pnet_netdev_event(struct notifier_block *this,
 711                                  unsigned long event, void *ptr)
 712 {
 713         struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
 714
 715         switch (event) {
 716         case NETDEV_REBOOT:
 717         case NETDEV_UNREGISTER:
 718                 smc_pnet_remove_by_ndev(event_dev);
 719                 return NOTIFY_OK;
 720         case NETDEV_REGISTER:
 721                 smc_pnet_add_by_ndev(event_dev);
 722                 return NOTIFY_OK;
 723         default:
 724                 return NOTIFY_DONE;
 725         }
 726 }
 727
 728 static struct notifier_block smc_netdev_notifier = {
 729         .notifier_call = smc_pnet_netdev_event
 730 };
 731
 732 /* init network namespace */
 733 int smc_pnet_net_init(struct net *net)
 734 {
 735         struct smc_net *sn = net_generic(net, smc_net_id);
 736         struct smc_pnettable *pnettable = &sn->pnettable;
 737
 738         INIT_LIST_HEAD(&pnettable->pnetlist);
 739         rwlock_init(&pnettable->lock);
 740
 741         return 0;
 742 }
 743
 744 int __init smc_pnet_init(void)
 745 {
 746         int rc;
 747
 748         rc = genl_register_family(&smc_pnet_nl_family);
 749         if (rc)
 750                 return rc;
 751         rc = register_netdevice_notifier(&smc_netdev_notifier);
 752         if (rc)
 753                 genl_unregister_family(&smc_pnet_nl_family);
 754         return rc;
 755 }
 756
 757 /* exit network namespace */
 758 void smc_pnet_net_exit(struct net *net)
 759 {
 760         /* flush pnet table */
 761         smc_pnet_remove_by_pnetid(net, NULL);
 762 }
 763
 764 void smc_pnet_exit(void)
 765 {
 766         unregister_netdevice_notifier(&smc_netdev_notifier);
 767         genl_unregister_family(&smc_pnet_nl_family);
 768 }
 769
 770 /* Determine one base device for stacked net devices.
 771  * If the lower device level contains more than one devices
 772  * (for instance with bonding slaves), just the first device
 773  * is used to reach a base device.
 774  */
 775 static struct net_device *pnet_find_base_ndev(struct net_device *ndev)
 776 {
 777         int i, nest_lvl;
 778
 779         rtnl_lock();
 780         nest_lvl = ndev->lower_level;
 781         for (i = 0; i < nest_lvl; i++) {
 782                 struct list_head *lower = &ndev->adj_list.lower;
 783
 784                 if (list_empty(lower))
 785                         break;
 786                 lower = lower->next;
 787                 ndev = netdev_lower_get_next(ndev, &lower);
 788         }
 789         rtnl_unlock();
 790         return ndev;
 791 }
 792
 793 static int smc_pnet_find_ndev_pnetid_by_table(struct net_device *ndev,
 794                                               u8 *pnetid)
 795 {
 796         struct smc_pnettable *pnettable;
 797         struct net *net = dev_net(ndev);
 798         struct smc_pnetentry *pnetelem;
 799         struct smc_net *sn;
 800         int rc = -ENOENT;
 801
 802         /* get pnettable for namespace */
 803         sn = net_generic(net, smc_net_id);
 804         pnettable = &sn->pnettable;
 805
 806         read_lock(&pnettable->lock);
 807         list_for_each_entry(pnetelem, &pnettable->pnetlist, list) {
 808                 if (pnetelem->type == SMC_PNET_ETH && ndev == pnetelem->ndev) {
 809                         /* get pnetid of netdev device */
 810                         memcpy(pnetid, pnetelem->pnet_name, SMC_MAX_PNETID_LEN);
 811                         rc = 0;
 812                         break;
 813                 }
 814         }
 815         read_unlock(&pnettable->lock);
 816         return rc;
 817 }
 818
 819 /* find a roce device for the given pnetid */
 820 static void _smc_pnet_find_roce_by_pnetid(u8 *pnet_id,
 821                                           struct smc_init_info *ini,
 822                                           struct smc_ib_device *known_dev)
 823 {
 824         struct smc_ib_device *ibdev;
 825         int i;
 826
 827         ini->ib_dev = NULL;
 828         spin_lock(&smc_ib_devices.lock);
 829         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 830                 if (ibdev == known_dev)
 831                         continue;
 832                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 833                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 834                                 continue;
 835                         if (smc_pnet_match(ibdev->pnetid[i - 1], pnet_id) &&
 836                             smc_ib_port_active(ibdev, i) &&
 837                             !test_bit(i - 1, ibdev->ports_going_away) &&
 838                             !smc_ib_determine_gid(ibdev, i, ini->vlan_id,
 839                                                   ini->ib_gid, NULL)) {
 840                                 ini->ib_dev = ibdev;
 841                                 ini->ib_port = i;
 842                                 goto out;
 843                         }
 844                 }
 845         }
 846 out:
 847         spin_unlock(&smc_ib_devices.lock);
 848 }
 849
 850 /* find alternate roce device with same pnet_id and vlan_id */
 851 void smc_pnet_find_alt_roce(struct smc_link_group *lgr,
 852                             struct smc_init_info *ini,
 853                             struct smc_ib_device *known_dev)
 854 {
 855         _smc_pnet_find_roce_by_pnetid(lgr->pnet_id, ini, known_dev);
 856 }
 857
 858 /* if handshake network device belongs to a roce device, return its
 859  * IB device and port
 860  */
 861 static void smc_pnet_find_rdma_dev(struct net_device *netdev,
 862                                    struct smc_init_info *ini)
 863 {
 864         struct smc_ib_device *ibdev;
 865
 866         spin_lock(&smc_ib_devices.lock);
 867         list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
 868                 struct net_device *ndev;
 869                 int i;
 870
 871                 for (i = 1; i <= SMC_MAX_PORTS; i++) {
 872                         if (!rdma_is_port_valid(ibdev->ibdev, i))
 873                                 continue;
 874                         if (!ibdev->ibdev->ops.get_netdev)
 875                                 continue;
 876                         ndev = ibdev->ibdev->ops.get_netdev(ibdev->ibdev, i);
 877                         if (!ndev)
 878                                 continue;
 879                         dev_put(ndev);
 880                         if (netdev == ndev &&
 881                             smc_ib_port_active(ibdev, i) &&
 882                             !test_bit(i - 1, ibdev->ports_going_away) &&
 883                             !smc_ib_determine_gid(ibdev, i, ini->vlan_id,
 884                                                   ini->ib_gid, NULL)) {
 885                                 ini->ib_dev = ibdev;
 886                                 ini->ib_port = i;
 887                                 break;
 888                         }
 889                 }
 890         }
 891         spin_unlock(&smc_ib_devices.lock);
 892 }
 893
 894 /* Determine the corresponding IB device port based on the hardware PNETID.
 895  * Searching stops at the first matching active IB device port with vlan_id
 896  * configured.
 897  * If nothing found, check pnetid table.
 898  * If nothing found, try to use handshake device
 899  */
 900 static void smc_pnet_find_roce_by_pnetid(struct net_device *ndev,
 901                                          struct smc_init_info *ini)
 902 {
 903         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 904
 905         ndev = pnet_find_base_ndev(ndev);
 906         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 907                                    ndev_pnetid) &&
 908             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid)) {
 909                 smc_pnet_find_rdma_dev(ndev, ini);
 910                 return; /* pnetid could not be determined */
 911         }
 912         _smc_pnet_find_roce_by_pnetid(ndev_pnetid, ini, NULL);
 913 }
 914
 915 static void smc_pnet_find_ism_by_pnetid(struct net_device *ndev,
 916                                         struct smc_init_info *ini)
 917 {
 918         u8 ndev_pnetid[SMC_MAX_PNETID_LEN];
 919         struct smcd_dev *ismdev;
 920
 921         ndev = pnet_find_base_ndev(ndev);
 922         if (smc_pnetid_by_dev_port(ndev->dev.parent, ndev->dev_port,
 923                                    ndev_pnetid) &&
 924             smc_pnet_find_ndev_pnetid_by_table(ndev, ndev_pnetid))
 925                 return; /* pnetid could not be determined */
 926
 927         spin_lock(&smcd_dev_list.lock);
 928         list_for_each_entry(ismdev, &smcd_dev_list.list, list) {
 929                 if (smc_pnet_match(ismdev->pnetid, ndev_pnetid) &&
 930                     !ismdev->going_away) {
 931                         ini->ism_dev = ismdev;
 932                         break;
 933                 }
 934         }
 935         spin_unlock(&smcd_dev_list.lock);
 936 }
 937
 938 /* PNET table analysis for a given sock:
 939  * determine ib_device and port belonging to used internal TCP socket
 940  * ethernet interface.
 941  */
 942 void smc_pnet_find_roce_resource(struct sock *sk, struct smc_init_info *ini)
 943 {
 944         struct dst_entry *dst = sk_dst_get(sk);
 945
 946         ini->ib_dev = NULL;
 947         ini->ib_port = 0;
 948         if (!dst)
 949                 goto out;
 950         if (!dst->dev)
 951                 goto out_rel;
 952
 953         smc_pnet_find_roce_by_pnetid(dst->dev, ini);
 954
 955 out_rel:
 956         dst_release(dst);
 957 out:
 958         return;
 959 }
 960
 961 void smc_pnet_find_ism_resource(struct sock *sk, struct smc_init_info *ini)
 962 {
 963         struct dst_entry *dst = sk_dst_get(sk);
 964
 965         ini->ism_dev = NULL;
 966         if (!dst)
 967                 goto out;
 968         if (!dst->dev)
 969                 goto out_rel;
 970
 971         smc_pnet_find_ism_by_pnetid(dst->dev, ini);
 972
 973 out_rel:
 974         dst_release(dst);
 975 out:
 976         return;
 977 }
 978
 979 /* Lookup and apply a pnet table entry to the given ib device.
 980  */
 981 int smc_pnetid_by_table_ib(struct smc_ib_device *smcibdev, u8 ib_port)
 982 {
 983         char *ib_name = smcibdev->ibdev->name;
 984         struct smc_pnettable *pnettable;
 985         struct smc_pnetentry *tmp_pe;
 986         struct smc_net *sn;
 987         int rc = -ENOENT;
 988
 989         /* get pnettable for init namespace */
 990         sn = net_generic(&init_net, smc_net_id);
 991         pnettable = &sn->pnettable;
 992
 993         read_lock(&pnettable->lock);
 994         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
 995                 if (tmp_pe->type == SMC_PNET_IB &&
 996                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX) &&
 997                     tmp_pe->ib_port == ib_port) {
 998                         smc_pnet_apply_ib(smcibdev, ib_port, tmp_pe->pnet_name);
 999                         rc = 0;
1000                         break;
1001                 }
1002         }
1003         read_unlock(&pnettable->lock);
1004
1005         return rc;
1006 }
1007
1008 /* Lookup and apply a pnet table entry to the given smcd device.
1009  */
1010 int smc_pnetid_by_table_smcd(struct smcd_dev *smcddev)
1011 {
1012         const char *ib_name = dev_name(&smcddev->dev);
1013         struct smc_pnettable *pnettable;
1014         struct smc_pnetentry *tmp_pe;
1015         struct smc_net *sn;
1016         int rc = -ENOENT;
1017
1018         /* get pnettable for init namespace */
1019         sn = net_generic(&init_net, smc_net_id);
1020         pnettable = &sn->pnettable;
1021
1022         read_lock(&pnettable->lock);
1023         list_for_each_entry(tmp_pe, &pnettable->pnetlist, list) {
1024                 if (tmp_pe->type == SMC_PNET_IB &&
1025                     !strncmp(tmp_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX)) {
1026                         smc_pnet_apply_smcd(smcddev, tmp_pe->pnet_name);
1027                         rc = 0;
1028                         break;
1029                 }
1030         }
1031         read_unlock(&pnettable->lock);
1032
1033         return rc;
1034 }